That's exactly the wrong way to think about it, and I'm surprised that so many devs think of it that way. We already have programs that works exactly like that (i.e. producing rich, complex output that would be many times the size of the input code + data if encoded raw): procedural generators. It's emergent complexity, not compression.